We conduct experiments on five subtasks over the FFHQ dataset with state-of-the-art baselines, and choose five most competitive baselines to tackle the fully-degraded Face Renovation task alongside HiFaceGAN.
| Degradation Type | Corresponding Task | Example Method |
|---|---|---|
| Downsampling | Super Resolution | ESRGAN |
| Mosaic | Face Hallucination | Super-FAN |
| Additive Noise | Denoising | WaveletCNN |
| Motion/Gaussian Blur | Deblurring | DeblurGANv2 |
| JPEG artifacts | Compression Artifact Removal | ARCNN |
| Mixed/Unknown | Face Renovation | HiFaceGAN |
Example of degraded images are showcased here:
And the quantitative performance are reported below:
